15/6/2019

Goal

Creation of value should be constrained only by our ability to generate good ideas.

  • Value stream : the sequence of activities an organization undertakes to deliver a good or service to a customer.
  • Flow of work : the work needed for a data driven feature to go from a hypothesis to operationalization.

But what IS DevOps

The OODA loop

Legendary fighter pilot and strategist John Boyd :
Having a tighter OODA loop than your competitors = Winning.

The OODA loop

Legendary fighter pilot and strategist John Boyd :
Having a tighter OODA loop than your competitors = Winning.

The core, chronic conflict and 2 cultures

a concept from The Goal by Eliyahu Goldratt

When two parts of an organization are driven by opposing incentive structures, outcomes for the organization are sub-optimal (or disastrous).

In IT :

  • Dev teams are incentivised to implement features to create business value
  • IT operations is incentivised to prevent outages

The core, chronic conflict and 2 cultures

For data science :

  • Data scientists are incentivised to frequently re-train and upgrade mathematical models
  • Dev teams will refuse to reimplement the same “feature” every 2 sprints

Why DataOps ?

Applying DevOps principles to extract maximal value from data science.

What we want to achieve :

Creation of value should be constrained only by our ability to generate good ideas.

  • Fast flow of work from a idea to value
    • fast deployments
    • close alignments
  • Strong feedback loops
    • transparent tracking
    • flow of information
  • Short cycle time for experimentation
    • minimise effort downstream of model deployment
    • minimise fear, maximize learning

Why is this hard ?

  • Three silos that make it challenging to create business value from data
    • Data science teams with many potential business problems to focus on
    • Development teams with long backlogs of feature requests
    • Business operations working “out there”
  • There are (at least) three unpredictable time scales
    • Time taken to develop a data driven solution to a business problem
    • Waiting time until a scrum team can prioritize the development of a new data driven feature
    • Time needed for the business ops changes to realize the benefits of a new feature

Okay fine, but How ?

  • Organization
    • team structure : data scientists, data engineeers and developers working together
    • “chapters” : closely interacting groups of people with a common skillset working in different scrum teams
  • Tooling
    • version control
    • docker
    • CI/CD pipelines
    • APIs
    • K8s

Okay fine, but How ?

Version control

Version control

  • The model “out there” can be traced back to a commit in the repo
  • Each commit can be traced back to a card in JIRA which contains the business justification for any changes made

Docker

API for your model

Useful resources

The take home message

  • Creation of value should be constrained only by our ability to generate good ideas.
  • Ever tighter OODA loop = winning

Thank you